DataSapien Methodologies
How we count, what we measure, and where the numbers come from. The figures on the DataSapien homepage are working estimates, sourced from external research where possible and from our own customer data where appropriate. Every claim on this page can be inspected. If you find something that needs sharpening, tell us.
Last updated: May 2026.
1.5 billion smartphones can now run generative AI on-device
The claim
Approximately 1.5 billion smartphones in active use today are capable of running generative AI models locally on the device — without sending data to a cloud service for inference.
This is a statement about capability, not current activation. The hardware, memory, and operating-system support are in place on these devices. The customer relationship is moving to where these devices live; the platforms that meet customers in that environment will compound the advantage.
The basis
“Capable of running generative AI on-device” is a specific technical threshold. We count a device only if it meets all four conditions:
- Silicon: a dedicated neural processing unit (NPU) with at least 30 TOPS of compute capacity, or equivalent GPU-accelerated AI capability
- Memory: at least 8 GB of RAM, sufficient to hold a small language model in active memory alongside the operating system and host application
- OS support: ships with vendor on-device AI runtime (Apple Intelligence, Google AICore via Android 14+, or Samsung Galaxy AI / OEM-equivalent runtime)
- In active use: counted as part of the global active smartphone installed base, not as a historical shipment number
Devices that meet some but not all of these conditions are excluded. The number is therefore a conservative floor, not a maximal claim.
The calculation
| Device class | Approximate active devices | Source basis |
|---|---|---|
| iPhone 15 Pro / Pro Max | ~120 million | Apple installed-base estimates derived from Counterpoint and Wave7 tracking, late 2025 |
| iPhone 16 series (16, 16 Plus, 16 Pro, 16 Pro Max, 16e) | ~280 million | Apple’s largest single-generation cohort to date |
| iPhone 17 series (17, 17 Pro, 17 Pro Max, Air) | ~80 million (partial cycle, FY26 to date) | Estimated from Apple’s quarterly reports |
| Samsung Galaxy S24 / S24+ / S24 Ultra | ~80 million | Counterpoint Galaxy tracking |
| Samsung Galaxy S25 series | ~120 million | Combined estimate from Counterpoint, IDC |
| Google Pixel 8, 9, 10 Pro / Pro XL series | ~60 million | Google does not disclose installed base; estimate triangulated from Statcounter, Sensor Tower |
| Snapdragon 8 Gen 3 / 8 Elite / 8 Gen 5 Android devices (excl. Samsung / Pixel) | ~480 million | Qualcomm Snapdragon device tracking + IDC |
| Dimensity 9300 / 9400 / 9500 Android devices | ~180 million | MediaTek + Counterpoint Android segment |
| Subtotal | ~1.4–1.5 billion |
Working figure rounded to 1.5 billion. The number is dynamic — every quarter adds roughly 80–120 million new capable devices as flagship and upper-midrange smartphones ship. Older devices age out of the installed base, but at a slower rate than new capable devices enter it.
What we excluded
We deliberately excluded:
- Smartphones with NPUs below 30 TOPS. Many Android devices ship with NPUs in the 5–25 TOPS range that can run small classifiers but not meaningful generative AI workloads.
- Devices with less than 8 GB of RAM. Memory is the binding constraint for on-device LLM inference, more than raw compute.
- Devices running cloud-routed AI features. Apple Intelligence routes some workloads to Private Cloud Compute; Google routes some workloads to Gemini cloud. These don’t count toward “on-device” capability even when the user-facing feature is similar.
- Tablets, laptops, wearables. The homepage figure is specifically about smartphones. The capable installed base across all device classes (including iPads with M-series silicon and AI-capable laptops) is materially larger but outside the scope of this claim.
Sources
- Counterpoint Research, Global Smartphone Installed Base Tracker, 2025 and 2026 updates
- IDC, Worldwide Mobile Phone Forecast, 2025
- Apple Inc., quarterly 10-Q filings (FY24–FY26)
- Google, AICore and Android AI documentation, 2025
- Samsung Electronics, Galaxy AI capability matrix, 2025
- Qualcomm, Snapdragon device tracking and partner shipment data, 2025
- MediaTek, Dimensity flagship device tracker, 2025
- TSMC, GenAI smartphone penetration forecasts, 2024–28
Caveats and confidence
This figure is a working estimate, not a precise count. The confidence interval is approximately ±200 million — i.e. the true number sits somewhere between 1.3 and 1.7 billion.
Three things would shift the number materially:
- Definition of “capable.” A stricter definition (e.g. requiring 12 GB RAM and 40+ TOPS) would reduce the count to around 800 million. A looser definition (any NPU plus 6 GB RAM) would raise it past 2.5 billion. We chose 30 TOPS and 8 GB as the threshold most aligned with “can run a 3-billion-parameter language model with acceptable latency.”
- Active installed base assumptions. Counterpoint’s installed-base data is updated quarterly. New flagship launches and replacement-cycle changes can shift the underlying numbers by 5–10% on a quarterly basis.
- OEM disclosures. Apple discloses unit sales but not active installed base. Google does not disclose Pixel installed base at all. The Android segment is therefore estimated rather than measured.
We re-validate this figure every six months against updated tracker data. The next review is scheduled for November 2026.
44× uplift in activation effectiveness when personalisation runs on the customer’s device
The claim
A controlled comparison across a recent DataSapien customer deployment showed personalisation activation rates 44× higher when the same offers were delivered via on-device, context-aware personalisation than when delivered via conventional cloud-based segmentation and broadcast.
The basis
The figure is derived from a single customer engagement, not a meta-analysis across multiple customers. It is a real result from a real deployment, not a projection or a synthetic benchmark. The customer has authorised the comparison to be used in our materials on an anonymised basis.
The comparison was structured as:
- Control arm: the customer’s existing personalisation stack — segmentation in their CDP, broadcast delivery via email and push, contextual targeting based on lifecycle stage and last-purchase data
- Treatment arm: the same offers, the same audience, the same redemption mechanics, but delivered via DataSapien’s on-device runtime, with personalisation decisions made on the device based on live context
Activation was measured as a successful customer interaction with the offer — defined as the customer engaging with the offer content within the redemption window.
The calculation
Across the eight-week comparison window:
- Control arm activation rate: ~0.18%
- Treatment arm activation rate: ~7.9%
- Ratio: 43.9× (rounded to 44× in marketing materials)
The 44× ratio is the headline number. The underlying activation rates are the more useful figures for benchmarking against other personalisation systems.
Why the lift is so large
Three structural factors compound:
- Timing. On-device personalisation can act on real-time context (location, time of day, current app state, recent interactions) without the latency and freshness penalty of cloud round-trips. Most cloud personalisation systems work with context that is hours or days old.
- Relevance. Live context allows offers to be matched to the customer’s actual current state, not their average historical state. The control arm’s segmentation reflects the customer as they were last week; the treatment arm reflects the customer as they are now.
- Permission. Customers responded to on-device offers at materially higher rates because the offers were delivered in a way that the customer had explicitly authorised, in a moment they perceived as theirs. Trust compounds with relevance.
A subset of the lift is explained by each factor individually; the combined effect is multiplicative, not additive.
What this number is and is not
It is:
- A real measured outcome from a real customer
- A point estimate, not a guarantee
- Specific to a particular customer, vertical, and offer type
It is not:
- A figure that will apply identically to every customer
- A range, distribution, or average across customers
- A projection of future performance
We expect lifts of similar magnitude in similar contexts (membership-led B2H categories, mobile-app-led customer relationships, real-time offer mechanics). We do not yet have enough deployments to publish a confidence interval. As more customer data accumulates, this section will be updated with a range and a sample size.
Sources
- DataSapien customer deployment data, FY25 (anonymised, authorised for external use)
- Internal experiment methodology documentation, retained for audit
- A more detailed technical derivation note is available on request to qualified partners and prospects
Caveats
- The 44× ratio is specific to one customer comparison. We use it because it is real, recent, and authorised. We do not claim it as an average across our customer base.
- The control arm was the customer’s existing system, which is a reasonable benchmark for the kind of buyer we typically talk to (a brand with an established CDP and broadcast marketing stack). It is not the best possible cloud-based system that could be constructed.
- The eight-week window captures activation, not long-term outcomes (retention, lifetime value, customer satisfaction). Those metrics are tracked separately and reported in customer case studies.
How we update this page
We update the figures on this page when:
- New external research changes the basis for a claim (e.g. Counterpoint publishes updated installed-base data)
- A new customer deployment provides validated data that complements or revises an existing claim
- The underlying definitions or methodology change
Each update is logged at the bottom of the relevant section. The current state of the page is always considered authoritative; older versions are retained internally for audit.
If you have questions about any figure, or you’d like to see additional working for a specific claim, please get in touch.
Last updated: May 2026. Next scheduled review: November 2026.